DiffusionGemma
by Google
Visit Website ↗

DiffusionGemma is Google's first family of diffusion language models, distilled from autoregressive Gemma models to generate text through denoising rather than token-by-token prediction. It supports parallel synthesis, flexible infilling, and controllable-length generation for next-generation AI research.

2
Total mentions
2
Articles
See website
Pricing
No
Free tier
← Back to registry
Articles mentioning DiffusionGemma (2)
01DiffusionGemma_Proves_Text_AI_Needs_No_Training_From_ScratchAug 9, 2026 02Google_Open_Model_DiffusionGemma_Text_From_NoiseJun 10, 2026
Key Features
Non-autoregressive generation via denoising LM process
Arbitrary text infilling and middle-out generation
Controllable output length and early-stop decoding
Available in 2B and 7B open-weight variants
Distilled from Gemma for strong world knowledge and reasoning
Pros & Cons
Pros
Open weights available under permissive Gemma license
Enables novel capabilities like infilling that causal LLMs lack
Parallel generation can be significantly faster than autoregressive decoding
Cons
Generation quality generally lags comparable autoregressive models
Requires custom inference pipelines and multiple denoising steps
Relatively new and not yet widely integrated into production tooling